A total of 661 samples were retrieved from the Gene Expression Omnibus (GEO) database, including 153 from healthy controls and 448 from patients with sepsis. The GSE13904, GSE26440, GSE28750, GSE95233, and GSE57065 datasets were obtained from the Affymetrix GPL570 platform (Affymetrix Human Genome U133 Plus 2.0 Array), while the GSE65682 dataset was derived from the Affymetrix GPL13667 platform (Affymetrix Human Genome U219 Array). Raw Affymetrix data were preprocessed using the robust multiarray averaging (RMA) algorithm implemented in the affy R package. A training set was constructed from GSE28750, GSE95233, and GSE57065 (GPL570 platform), and validation set 1 was generated by merging GSE13904 and GSE26440. Validation set 2 was constructed from GSE65682 (GPL13667 platform). All included samples were derived from peripheral blood.

To examine the expression profile of FTO across immune cell populations, single-cell RNA sequencing (scRNA-seq) data from the GSE249975 dataset were analysed.

To systematically identify molecular signatures associated with sepsis, we employed a broad panel of machine learning algorithms covering linear, nonlinear, and ensemble learning paradigms. The linear methods included logistic regression, partial least squares (PLS) regression, least absolute shrinkage and selection operator (LASSO), and linear discriminant analysis (LDA). The nonlinear methods included support vector machines (SVMs) with a kernel function, k-nearest neighbour (k-NN) models, artificial neural networks (ANNs), and naïve Bayes methods. The ensemble approaches included random forest models, the generalized boosting method, the gradient boosting machine (GBM), adaptive boosting (AdaBoost), categorical boosting (CatBoost), and the light gradient boosting machine (LightGBM). Both the GBM and generalized boosting method were evaluated, as they represent widely used but technically distinct implementations of gradient boosting in biomedical research.

Model performance was assessed using receiver operating characteristic (ROC) curves, precision–recall (PR) curves, and the area under the curve (AUC). Moreover, sensitivity, specificity, and calibration were evaluated. The AUC was considered the primary metric. All analyses were conducted in R (version 4.3.2). The caret package was used for model training and cross-validation, while pROC, kernelshap, rmda, catboost, lightgbm, and DALEX were applied for model evaluation, interpretation, and visualization.

m6A2Target (http://m6a2target.canceromics.org/) is a comprehensive database that provides information on the target genes of writers, erasers, and readers (WERs). It integrates high-confidence targets validated by low-throughput experiments as well as potential targets supported by binding evidence from high-throughput sequencing or inferred from WER perturbation followed by high-throughput assays (21). The putative gene targets of key m⁶A regulators associated with disease diagnosis were identified using m6A2Target. Functional annotation of biological processes (BPs) was conducted using the GOplot package, while functional enrichment analyses were performed with the ClusterProfiler package (version 4.10.1), a widely used tool for the interpretation of omics data.

Single-cell RNA sequencing data were obtained from a publicly available dataset derived from peritoneal immune cells of CLP-induced septic mice. The dataset was analyzed as provided by the original study. Raw 10× Genomics data were imported and converted into Seurat objects using the Seurat package (version 4.4.0). Quality control, normalization, dimensionality reduction, and clustering of peritoneal cell populations were subsequently performed according to standard Seurat workflows. Cell subsets were annotated based on reference transcriptomic datasets provided by the celldex package (version 1.12.0).

In total, 12 patients with sepsis and 12 healthy controls were included. The diagnosis of sepsis was established according to the Third International Consensus Definitions for Sepsis and Sepsis Shock (Sepsis-3) (22). After receiving written informed consent, peripheral blood samples were collected and PBMCs and neutrophils were promptly separated. The experiment has been approved by the Ethics Committee of China-Japan Friendship Hospital (approval number 2024-KY-417-1).

Polymicrobial sepsis was established in mice using the CLP procedure, as previously described (23) with minor modifications. Mice were anesthetized with 1.5% isoflurane delivered at 2 L/min. Following standard abdominal disinfection, a midline incision of approximately 1.5–2 cm was made to access the peritoneal cavity. The cecum was carefully isolated from the mesentery, and the distal portion was ligated with sterile silk at approximately halfway along its length. Two punctures were performed through the ligated portion using a 21-gauge needle, and a small amount of fecal material was gently extruded to ensure patency. The abdominal wall and skin were then closed in layers. Sham-operated controls underwent the same procedure except that the cecum was neither ligated nor punctured.

RAW264.7 cells were chosen for macrophage-related experiments. To suppress the expression of FTO, Fto-specific siRNA (siFTO) and negative control siRNA (siNC) were chemically synthesized by D-Nano Therapeutics (China). Transfection was carried out using the CALNP™ RNAi transfection reagent (D-Nano Therapeutics, China) following the manufacturer’s protocol. For rescue assays, transfected cells were treated with 20 µM PI3K agonist 740 Y-P (HY-P0175, MCE). The siRNA sequences are listed in Supplementary Table 1.

Primary neutrophils were isolated from mouse bone marrow as previously described and cultured in RPMI-1640 with 10% FBS (24). Cells were stimulated with the FTO inhibitor FB23-2 (HY127103, MCE) and/or 10 µM PI3K inhibitor LY294002 (HY10108, MCE) prior to downstream assays.

RNA isolation and real-time quantitative polymerase chain reaction (RT–qPCR) analysis were performed as previously described (25). GAPDH was used as the internal control. The primer sequences employed in this study are listed in Supplementary Table 2.

Protein extraction and Western blot analysis were performed as previously described (25). The primary antibody used was anti-FTO (EPR24440-12, 1:50 dilution; abcam), p-PI3K (T40116M, 1:1000; Abmart), PI3K (T40115M, 1:1000; Abmart), p-AKT (#4060, 1:1000; CST), AKT (#4685, 1:1000; CST), with GAPDH and β-actin serving as the internal control. Band intensities on autoradiograms were quantified by densitometric analysis using Quantity One software (Bio-Rad).

Following stimulation with LPS (200 ng/mL) for 24 h, the cells were harvested and washed with PBS. Surface staining was performed at 4 °C for 30 min using PE-conjugated anti-F4/80 (cat. 123109), FITC-conjugated anti-CD11b (cat. 101205), and Brilliant Violet 421-conjugated anti-CD86 (cat. 105031). Flow cytometry was conducted on a Beckman instrument, and the data were analysed using FACSDeva software.

After transfection and stimulation, the cell culture supernatants were collected, and the level of TNF-α was quantified using a mouse TNF-α ELISA kit (E-HSEL-M0009; Elabscience, China) according to the manufacturer’s protocol.

Intracellular reactive oxygen species (ROS) detection generation was assessed using DCFH-DA (Reactive Oxygen Species Assay Kit, Solarbio, cat. CA1410) according to the manufacturer’s protocol. Following treatment, cells were incubated with 1:750 diluted DCFH-DA for 30 min at 37 °C in the dark, washed twice with serum-free medium, and counterstained with Hoechst 33342 (Solarbio). Fluorescence was recorded at 488/525 nm (excitation/emission), and ROS levels were expressed as mean fluorescence intensity.

Immunofluorescence staining was performed to evaluate FTO expression and its spatial association with immune cell populations in intestinal tissues. Tissue sections were fixed with 4% paraformaldehyde for 20 min, permeabilized with 0.3% Triton X-100 for 10 min, and blocked with 5% bovine serum albumin for 1 h at room temperature. Sections were then incubated overnight at 4 °C with primary antibodies against FTO (EPR24440-12, 1:50 dilution; abcam), F4/80 (ab04467, 1:200 dilution; abcam) or Ly6G (cat. 127609, 1:200 dilution; BioLegend). After washing, sections were incubated with Alexa Fluor 488- or 594-conjugated secondary antibodies (1:500) for 1 h at room temperature in the dark, followed by nuclear counterstaining with DAPI. Images were acquired using a Zeiss laser scanning confocal microscope under identical settings for all groups.

Statistical analyses were performed using GraphPad Prism (GraphPad Software, San Diego, CA) and R software (version 4.3.2). Comparisons between two groups were conducted using Student’s t-test. For experiments involving more than two groups, one-way analysis of variance (ANOVA) followed by Tukey’s post hoc multiple-comparison test was applied. Data are presented as the mean ± standard error of the mean (SEM), and P < 0.05 was considered statistically significant.

Given the pivotal role of m⁶A methylation regulators in shaping inflammation and immunity, we systematically evaluated the diagnostic potential of 24 m⁶A regulators for sepsis using publicly available datasets. On the basis of the expression profiles of these regulators, we constructed diagnostic models employing 14 distinct machine learning algorithms, namely, logistic regression, PLS, LASSO, LDA, SVM_Kernel, k-NN, ANN, naïve Bayes, random forest, generalized boosting, GBM, AdaBoost, CatBoost, and LightGBM algorithms. Rigorous cross-validation was conducted on the basis of the training set and two independent validation cohorts. The cross-validation performance in the training set is summarized in Supplementary Table 3, with all the models achieving AUC values greater than 0.9 (Figure 2A). To further place these diagnostic findings into a broader regulatory context, we conducted integrative pathway-level analyses across multiple independent sepsis transcriptomic datasets. Using ssGSEA, we performed a systematic comparison of m⁶A RNA modification versus other RNA modification and epigenetic regulatory pathways, including m1A, m5C, and m7G RNA modifications, as well as DNA methylation. Notably, m6A-related regulatory pathways exhibited the most consistent and pronounced dysregulation between control and sepsis samples across datasets (Supplementary Figure 1), thereby prioritizing m⁶A modification as a key regulatory mechanism for subsequent mechanistic investigation in this study.

To compare the diagnostic performance across algorithms, we next assessed their classification accuracy in the validation datasets. In validation set 1 (Figure 2B), the ANN model achieved the highest AUC (0.83), while logistic regression and k-NN had the lowest AUC values (~0.50). However, the specificity in validation set 1 was nearly zero (Supplementary Table 4), suggesting potential overdiagnosis; therefore, subsequent analyses focused mainly on validation set 2. In validation set 2 (Figure 2C), all the models achieved AUC values greater than 0.70, with LASSO reaching the highest AUC (0.99). Models including the GBM, logistic regression, PLS, generalized boosting, ANN, LDA, LASSO, AdaBoost, and LightGBM models maintained balanced accuracy, sensitivity, and specificity above 0.60, indicating satisfactory generalizability (Table 1).

In the residual boxplots (Figure 2D), compared with the other models, the LASSO, GBM, LightGBM, and generalized boosting models resulted in smaller residuals, indicating better generalization performance. Conversely, models such as PLS, logistic regression and ANN showed higher residuals, suggesting potential overfitting and limited robustness on unseen data. Decision curve analysis (Figure 2E) further demonstrated that the LASSO model consistently provided greater net clinical benefit than the other models across threshold probabilities between 0.5 and 1. Confusion matrices (Figure 2F) confirmed the robustness of the LASSO model, with accuracies of 98.5% in the training set and 97.6% in validation set 2. According to the clinical impact curves (Figure 2G), the LASSO model maintained stable predictive performance across both the training and validation cohorts. Collectively, these results highlight that, particularly for validation set 2, the LASSO model has strong potential for clinical application in the blood-based diagnosis of sepsis.

Using these machine learning approaches, different candidate biomarkers were identified across the validation sets. Given that the optimal models differed between the two validation cohorts, we calculated the mean |SHAP| values of 24 m⁶A regulators to evaluate their relative contributions to model construction. In validation set 1, the ANN model highlighted FTO as the top contributor, with a mean |SHAP| value of 0.189 (Figure 3A). In validation set 2, we focused on three models with superior diagnostic performance—LASSO, LightGBM, and CatBoost—all of which consistently ranked FTO as the most important feature, with mean |SHAP| values of 0.118 (Figure 3B), 0.285 (Figure 3C), and 1.524 (Figure 3D), respectively. Furthermore, analysis of the training set revealed significant differential expression of several m⁶A regulators between sepsis patients and healthy controls, among which FTO displayed one of the most pronounced differences (Figure 3E). Collectively, these findings underscore the central role of FTO in diagnostic modelling for sepsis.

To further explore the potential roles of m⁶A regulators in sepsis and to investigate the pathways involved, we used the m6A2Target database to predict downstream targets. For FTO, 105 experimentally validated targets and 23,943 predicted targets were identified (Supplementary Table 5). On the basis of the predicted targets, BPs enrichment analysis was performed using the ClusterProfiler package, revealing that FTO-related genes are involved mainly in apoptosis, autophagy, oxidative stress, innate immunity, and cell cycle regulation, all of which are critically involved in the pathogenesis of sepsis (Figure 3F). KEGG pathway enrichment analysis further demonstrated that the predicted FTO targets were significantly enriched in key signalling pathways, including the PI3K-Akt signalling pathway, MAPK signalling pathway, Ras signalling pathway, cell cycle, Hippo signalling pathway, autophagy, TGF-beta signalling pathway, apoptosis, p53 signalling pathway, and HIF-1 signalling pathway (Figure 3G). To better illustrate the regulatory relationships, a Cytoscape-based network was constructed, which revealed that FTO was closely connected with several hub targets, such as MC4R, ASB2, MYC, PPARG, and FOXO1 (Figure 3H).

Accumulating evidence has demonstrated that m⁶A modification serves as a critical regulatory factor in immune responses and inflammatory networks. To elucidate the immunological features associated with sepsis and FTO expression, we first evaluated the immune scores and immune cell fractions. In the training cohort, neutrophils accounted for the greatest proportion of infiltrating cells, followed by monocytes and M0 macrophages (Figure 4A). Comparative analysis revealed significant differences in the proportions of multiple immune cell subsets, including neutrophils, monocytes, M0 macrophages, CD4+ naive T cells, resting NK cells, M2 macrophages, M1 macrophages, resting mast cells, gamma delta T cells, plasma cells, naive B cells, memory B cells, CD8+ T cells, resting CD4+ memory T cells, follicular helper T cells, activated resting NK cells, resting dendritic cells, activated mast cells, and eosinophils (Figure 4B), between sepsis patients and healthy controls. Among these subsets, neutrophils and M1 macrophages are of particular interest because of their high relevance to sepsis pathophysiology (26–28). The proportion of neutrophils was significantly greater in sepsis patients than in controls and was negatively correlated with FTO expression (Figure 4C). Similarly, the proportion of M1 macrophages was markedly greater in sepsis patients and correlated with FTO levels (Figure 4D), suggesting the potential involvement of FTO in M1 macrophage activation.

To further characterize the relationship between FTO expression and neutrophil/macrophage populations, we performed scRNA-seq analysis using the GSE249975 dataset. Uniform manifold approximation and projection (UMAP) revealed 18 distinct cellular clusters (Figure 4E). On the basis of SingleR annotation, six major immune cell types were defined: neutrophils, macrophages, B cells, T cells, dendritic cells (DCs), and natural killer (NK) cells (Figures 4F, G). Notably, FTO was coexpressed with canonical neutrophil (Ly6g) and macrophage (Cd68) markers (Figure 4H). Cell type-specific analysis further revealed that FTO was predominantly expressed in neutrophils and macrophages (Figure 4I). When FTO expression was compared between the control and sepsis groups, we observed significant downregulation of FTO in both neutrophils and macrophages from sepsis patients (Figures 4J, K). These findings highlight the close association of FTO with innate immune cells and suggest its potential involvement in the dysregulated immune responses characteristic of sepsis.

To validate the clinical relevance of FTO dysregulation, we first examined its expression in PBMCs isolated from sepsis patients and healthy controls. RT–qPCR analysis demonstrated that FTO mRNA levels were significantly reduced in PBMCs from sepsis patients compared with those from healthy individuals (Figure 5A). We next investigated FTO expression in intestinal tissue from mice subjected to CLP–induced polymicrobial sepsis and sham-operated controls. Immunofluorescence staining revealed that FTO signals partially colocalized with F4/80⁺ macrophages in the intestinal mucosa, indicating that FTO is expressed in intestinal macrophage populations (Figure 5B). To further assess alterations in macrophage-associated FTO expression under septic conditions, the integrated fluorescence intensity of FTO within F4/80⁺ macrophages was quantitatively analyzed. As shown in Figure 5C, FTO in F4/80⁺ macrophages was significantly decreased in CLP mice compared with sham controls. Consistent with the in vivo findings, in vitro stimulation of RAW264.7 macrophages with LPS resulted in a marked downregulation of FTO mRNA expression at 4, 8, 12, and 24 h (Figure 5D). A corresponding reduction in FTO protein levels was observed after 24 h of LPS treatment (Figure 5E). To determine the functional role of FTO in macrophage activation, FTO expression was silenced in RAW264.7 cells using specific siRNAs. The knockdown efficiency was confirmed by RT–qPCR and Western blotting (Supplementary Figure 2A, B). To dissect the molecular mechanism by which FTO modulates macrophage activation, we focused on the PI3K/AKT signaling pathway based on KEGG pathway enrichment analysis. Western blot results showed that while LPS stimulation induced PI3K and AKT phosphorylation, siFTO significantly suppressed the levels of p-PI3K and p-AKT compared to the LPS control group (Figure 5F). This suggests that FTO is required for the sustainment of PI3K/AKT signaling during inflammatory responses. To determining whether the inhibition of PI3K/AKT signaling is responsible for the exacerbated inflammatory phenotype observed in FTO-deficient macrophages, we performed a rescue experiment using a specific PI3K agonist. RAW264.7 cells were transfected with siFTO and subsequently treated with LPS in the presence or absence of the PI3K agonist. As expected, the PI3K agonist effectively restored the phosphorylation levels of PI3K and AKT in siFTO-transfected cells (Figure 5F). We then evaluated the inflammatory phenotype. Flow cytometric analysis revealed that FTO silencing significantly increased the mean fluorescence intensity of the M1 marker CD86 compared with LPS treatment alone. Notably, treatment with the PI3K agonist significantly attenuated this siFTO-induced upregulation of CD86 (Figure 5G). Consistently, RT–qPCR and ELISA analyses demonstrated that the elevated TNF-α mRNA expression and protein secretion caused by FTO knockdown were markedly reversed by the restoration of PI3K/AKT signaling (Figures 5H, I). Furthermore, the enhanced intracellular ROS production observed in the siFTO group was significantly mitigated by the PI3K agonist (Figure 5J). Collectively, these data provide rigorous evidence that FTO prevents excessive macrophage inflammation, at least in part, by maintaining PI3K/AKT signaling axis.

To validate the clinical relevance of FTO dysregulation in neutrophils, we first examined its expression in neutrophils isolated from sepsis patients and healthy controls. RT–qPCR analysis revealed that FTO mRNA levels were significantly reduced in neutrophils from sepsis patients compared with those from healthy individuals (Figure 6A). We next investigated FTO expression in intestinal tissues from mice subjected to CLP–induced polymicrobial sepsis and sham-operated controls. Immunofluorescence staining demonstrated that FTO signals partially colocalized with Ly6G⁺ neutrophils in the intestinal mucosa, indicating that FTO is expressed in intestinal neutrophils (Figure 6B). To further assess alterations in neutrophil-associated FTO expression under septic conditions, the integrated fluorescence intensity of FTO within Ly6G⁺ neutrophils was quantitatively analyzed. As shown in Figure 6C, FTO expression in Ly6G⁺ neutrophils was significantly decreased in CLP mice compared with sham controls. In vitro, LPS stimulation induced dynamic changes in FTO expression in neutrophils. FTO mRNA levels were markedly downregulated at 4 and 8 h, followed by a gradual increase at 12 and 24 h after LPS treatment (Figure 6D). Consistent changes in FTO protein expression were observed at 8 h (Figure 6E). To further explore the functional relevance of FTO in neutrophils, the selective FTO inhibitor FB23–2 was applied to suppress FTO activity. RT–qPCR analysis showed that FB23–2 treatment significantly increased the mRNA expression of proinflammatory mediators, including TNF-α, IL-1β, CXCL2, CXCL3, and CXCL5 (Figures 6F–J). Based on KEGG pathway enrichment analysis, the PI3K/AKT signaling pathway was further examined. Pharmacological inhibition of PI3K with LY294002 markedly reversed the FB23-2–induced upregulation of inflammatory gene expression (Figure 6K). Western blot analysis demonstrated that LPS stimulation increased the phosphorylation levels of PI3K and AKT in neutrophils compared with untreated controls. Notably, FB23–2 treatment further enhanced the phosphorylation of PI3K and AKT without significantly affecting their total protein levels, whereas PI3K inhibition restored their phosphorylation status (Figure 6K). These results suggest that FTO dysregulation in neutrophils is associated with altered PI3K/AKT signaling during sepsis.